Lucie Fontaine from INRIA Bordeaux is visiting the Machine Learning group. Her research focuses on predictive coding models of hippocampus. She will present highlights from her latest work in this area.
Semantic and episodic memories in the neocortex and the hippocampus
The hippocampus and the neocortex can be seen as two Complementary Learning Systems: the former learns the specifics of individual episodes (episodic memory) whereas the latter extracts generalities across episodes (semantic memory). In the hippocampus, sparse, pattern-separated representations allow episodic memories to be stored quickly and to be recalled without interference. In the neocortex, dense, overlapping representations allow semantic memory to be learned slowly and to support generalization. After storage, the hippocampus replays episodic memories back to the neocortex to gradually transform them into semantic memory.
However, recent theoretical and empirical results led us to question this division. On the one hand, predictive coding networks, which are biologically inspired models of the visual cortex, have been shown to demonstrate hippocampus-like abilities on auto-associative memory tasks. On the other hand, the hippocampus has been shown to not only replay previous episodes, but also generate novel ones, supporting cognitive functions such as imagination and planning, in addition to memory consolidation. This evidence corroborates the recent theory that the hippocampus may itself contain complementary learning systems.
We propose a predictive coding model of hippocampal-cortical interactions, modelling the hippocampus implicitly, to investigate the ability of the neocortex to perform auto-associative memory and to support memory replay. Our results suggest that the neocortex cannot recall the specifics of individual experiences, but contributes to episodic memory by allowing episodic memories replayed by the hippocampus to be reinstantiated. We also present a preliminary model of the hippocampus, showing how its complementary learning pathways could support auto-associative memory and generative replay. With this combined neocortex-hippocampus model, we contribute to a better understanding of the mechanisms underlying the various cognitive functions supported by the hippocampus.
